Abstract: To reconstruct 2-D defect profile from magnetic flux leakage (MFL) signals, a dual wavelet neural network iteration model, including a forward model and an inverse model, based on radial wavelet basis function neural network was proposed. It iteratively adjusts the weights of the inverse network to minimize the error between the measured and predicted MFL signals. The network can be trained respectively by the same training samples from measurement and FEM calculation. To improve the network’s adaptability and accuracy, a novel training algorithm was proposed. Firstly, confirm the optimal number of layers, and then update the weights based on the conjugate gradient algorithm. The reconstruction results in different resolutions and SNRs indicate that the method is rapid, accurate and robust, and it is effective and feasible for reconstruction of 2-D defects comparing with other approaches.